Do you mean fine tuning (say via LoRA) or working with an "open weight" model or are you talking about true "open source" models. If it is the latter, I struggle to understand how a circa GPT3 model requiring millions and millions to train would be a good use of resources over one of the former.
Great clarifying question; sorry I should have been more clear.
I mean working with open weights and in the future, creating their own open source models from the ground up. At the moment, that's not reasonable, but I'm leaving the door open for when/if there's a break through in training that greatly reduces the costs of training new models.
Right now, LoRA can give amazing results with small models.
Do you mean fine tuning (say via LoRA) or working with an "open weight" model or are you talking about true "open source" models. If it is the latter, I struggle to understand how a circa GPT3 model requiring millions and millions to train would be a good use of resources over one of the former.
Great clarifying question; sorry I should have been more clear.
I mean working with open weights and in the future, creating their own open source models from the ground up. At the moment, that's not reasonable, but I'm leaving the door open for when/if there's a break through in training that greatly reduces the costs of training new models.
Right now, LoRA can give amazing results with small models.
I certainly hope so, DeepSeek has been giving out more of the code behind their systems
Likewise, I this direction gives engineers a way also to fight back with all the hackers/scammers. It'll create so many new companies as well.